A Scheme for Visual Feature based Image Indexing

1
A Scheme for Visual Feature based Image Indexing

• HongJiang Zhang and Di Zhong
• SPIE Conf. on Storage and Retrieval for Image and
  Video Databases
• Feb 1995
• Presented by Vibhore Vardhan

2
Motivation

• Digital images need to be manipulated and managed
  as images
• Retrieve visual data based on visual content of
  image
• E.g. IBMs QBIC system
• More emphasis on deriving visual features such as
  color, texture
• Need an effective indexing scheme utilizing these
  features
• Necessary to browse large image databases

3
Multi-dimensional Index

• Pre-computed visual features for each item in
  database
• Key attribute for an item is a feature vector
• Search based on similarities between feature
  vectors
• Three popular approaches to multi-dimensional
  indexing
• R tree, Linear quadtrees, and Grid files
• But they assume the following
• Distance Euclidean distance of points in
  feature space
• Dimensionality of feature space is low
• Efficient filter allows false positives, but no
  false dismissals
• Lower dimensional feature space or narrow search
  space

4
Tree Indexing by Abstraction and Classification

• For given attribute Aj, identify and label all
  objects that share Aj
• Identical value or a certain range al
• Objects with same label are clustered to form an
  abstraction
• Abstractions are represented as nodes in the
  index tree
• Apply these abstraction operations recursively to
  reach root node
• Automate using Self-Organization feature Maps
  (SOM)
• Unsupervised learning based on a grid of
  artificial neurons
• Weights are adapted to match input vectors in a
  training set
• First described by Teuvo Kohonen

5
Architecture of Self-Organization Map
of nodes gt of possible classes
associated weight
image feature vector

• Two layer SOM mapping from input data in Rn
  onto a 2-d grid
• All ref. vectors compared with 1 input vector
  according to metrics
• Select best matching node in the map, update
  neighbors
• After several iterations, SOM adapts to input

6
Hierarchical SOM

• Modify SOM to meet certain properties
• Constructs an index tree which forms similarity
  space of feature data
• First form bottom level L through learning
• Each node in L represents a group of image which
  are similar
• Higher levels are created by applying clustering
  and projection

7
Results Texture Features based Index
Iconic map of Brodatz texture database
8
Results Texture Features based Index
9
Results Texture Features based Index

• Texture feature set model (20 dimension feature
  vector)
• Multi-resolution simultaneous auto-regressive
  (MRSAR)
• Combined MRSAR with coarseness features and gray
  histograms
• Improves accuracy feature vector size goes up
  to 30
• Evaluated on Brodatz texture database
• 112 classes of images (512x512 8-bits)
• 9 subimages (128x128 8-bits) in each class
• Retrieval rate number of retrieved subimages
  (same class as query)
• number of
  retrieval neighbors

10
Results Texture Features based Index

• MRSAR does as well as global search for 9-nearest
  neighbor searching
• 5x speed improvement
• Adding coarseness and histogram improves accuracy
  by 6

11
Results Color Histogram based Index
Color histogram of images as feature vector for
317 images 106 images classified into 7
categories, rest act as noise
Compared indexing using color histograms in 3
color spaces Results for RGB space with 10
neighbors
Retrieval accuracy similar to global search, but
faster LUV color space gave the highest retrieval
rate
12
Conclusion

• Initial work in developing an effective indexing
  scheme
• Feature vector for images have high dimensions
• Not suitable for traditional indexing approaches
• Implements hierarchical SOM
• Results show good accuracy and speedups

13
DEMO

• PicSOM